Commission proposes ‘artificial intelligence factories’ in bid to give start-ups access to supercomputers

25 Jan 2024 | News

The prohibitive cost of computing is a big barrier for AI SMEs trying to enter the market. A new Commission proposal aims to ease the way, but questions over access, budget and an ethics requirement remain

The European Commission has unveiled a plan to create ‘AI factories’ across the bloc to give artificial intelligence start-ups access to supercomputers on which to build their own models, rather than relying on those created by US tech giants.

However, there are questions over whether this latest initiative has enough funding to stop the EU falling even further behind the US and China on AI.

“Computing infrastructure is a prohibitively expensive component of AI and this creates high barriers for start-ups to enter the market,” said Sebastiano Toffaletti, secretary general of the European Digital SME Alliance. “That's why the Commission's proposal is so important. By providing computing as a public good it allows more SMEs to build their own models rather than rely on big tech products,” he said.

The plan is to change the mandate of EuroHPC, the joint undertaking that plans and commissions European supercomputers, to give it a much stronger focus on AI.

The idea of opening up supercomputers to start-ups was floated in Commission president Ursula von der Leyen’s State of the Union address in September last year.

Following this, in November the Commission launched an AI grand challenge, under which four start-ups will get access to a supercomputer.

This week’s AI factories announcement goes further. EuroHPC will set up these facilities, commissioning new or upgrading existing, computers and pairing them with data centres, support services and access to expertise. These will “serve as a one-stop shop for Europe’s AI start-ups, enabling them to develop the most advanced AI models and industrial applications,” said internal market commissioner Thierry Breton in a statement.

There is no word yet on how many AI factories there will be, but access will be “free of charge for applications related to publicly funded research and innovation activities," a Commission spokesman said.

The machines could be adapted to both AI and scientific research, like the current EU supercomputers, or they could be targeted at AI and “not as suitable for science,” said EuroHPC director Anders Jensen.

Since von der Leyen’s speech last September, EuroHPC has changed its access policy and an update is due to be published in the coming days.

This will create a separate category for AI in the peer review process, to allow AI projects to be better compared to each other, and to make the application process, designed for scientific projects, better suited to industrial users.

There will be an ethics requirement for start-ups using the factories. “Only proposals for developing trustworthy and ethical artificial intelligence models, systems and applications that are in line with EU values shall be eligible for access,” according to EuroHPC’s amended mandate.

The Commission did not give further details on how this will work. “After assessing the validity and technical feasibility of the request, the EuroHPC grants access based on available capacity,” said the Commission spokesman. “Prioritisation of the requests is thus done for the moment on quality and pertinence. There is currently no prioritisation mechanism that uses other criteria, such as strategic or political arguments.”

EuroHPC expects the upcoming AI Act to provide more clarity. “My hope and expectation is we will not be left having to make the political judgement on what this means,” Jensen said.

There is no extra funding for the AI factories. Instead there will be a “redeployment” of resources from other EuroHPC projects, according to the details of the plan.

As with other EuroHPC supercomputers, the EU will cover 50% of the acquisition and deployment costs, and the country or countries in the hosting consortium will cover the rest.

It could prove controversial among scientists if money is siphoned off from research-focused supercomputers in favour of AI-tuned machines.

But Jensen argues this is not a problem, as the original regulation provided a “very generous budget” for supercomputers. Allowing investments in AI-dedicated machines will offer “a bit more freedom on how we spend the money that was already foreseen,” he said.

“I only hope that [EuroHPC] recognises the exiting potential in present and planned investments,” said Per Öster, director of the Advanced Computing Facility at CSC, the Finnish research centre that hosts Lumi, one of the EU’s leading supercomputers.

Lumi is acting as an AI factory right now, he said, and has already produced several large language models.

There has been concern that EuroHPC machines are not particularly suitable for AI work, but Öster says this is a misconception. Plus, dedicated AI capacity is a “very costly thing to do,” he said.

Toffaletti welcomed the pivot to AI-focused supercomputers but said the budget is too small. “It can only work if it leverages investments from the private sector,” he said.

For Holger Hoos, a founder of the Confederation of Laboratories for AI Research in Europe (CLAIRE), the plan risks spreading money and computing power too thinly across the bloc. “The AI factories could be useful, if set up and operated in close collaboration between the EuroHPC and AI communities, but they do not address the key need for focus and critical mass in AI research and innovation in Europe,” he said.

Instead, Hoos and other members of CLAIRE have argued for a much more concentrated approach, and for the EU to build what they call a CERN for AI.

Wider efforts

Beyond the AI factories initiative, EuroHPC supercomputers are already looking to give start-ups access. For example, the Jupiter supercomputer, to be based at the Jülich Research Centre near Cologne, should be opened up to start-ups, said Thomas Lippert, head of the centre. 

There is potential hurdle in the fact that EuroHPC supercomputers are half-funded by the member state in which they are located.

In the case of Jupiter, Germany might need to change its national regulations to allow companies to use Jupiter without being required to make their results public, Lippert noted.

In Finland, 20% of the Finnish share of Lumi’s supercomputing capacity is open for use by industry, Öster said. Access is free if results are openly published, or if companies collaborate with academia. Otherwise, Lumi charges for use.

Public versus private

The AI factories plan suggests the Commission is placing its hope in private firms to keep the EU competitive in AI.

Hoos, however, has argued this is misplaced, and the EU needs to publicly fund a Manhattan Project-style effort to build leading edge models.

Speaking to Science|Business after an AI summit between Germany, France and Poland earlier this week, German research minister Bettina Stark-Watzinger poured cold water on this approach.

“No state, or association of states, can match the investments made by large corporations like Microsoft or Google with public investments,” she said. “Moreover, this is not our economic model. Our economic success is built upon the ideas and initiative of entrepreneurs and private actors. We must create the right framework so that these ideas can be implemented.”

But Stark-Watzinger is optimistic about the EU’s position as an AI leader. The bloc has great strengths in basic research, image and language processing, AI applications in production and industry, logistics, business-to-business services, autonomous systems, and robotics, especially collaborative robotics, she said.

Never miss an update from Science|Business:   Newsletter sign-up